* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Chapter 3 - Department of Computer and Information Science and
Multiprotocol Label Switching wikipedia , lookup
Wake-on-LAN wikipedia , lookup
Cracking of wireless networks wikipedia , lookup
Asynchronous Transfer Mode wikipedia , lookup
Deep packet inspection wikipedia , lookup
Internet protocol suite wikipedia , lookup
UniPro protocol stack wikipedia , lookup
Recursive InterNetwork Architecture (RINA) wikipedia , lookup
Modeling & Analysis Mathematical Modeling: probability theory queuing theory application to network models Simulation: topology models traffic models dynamic models/failure models protocol models Transport Layer 3-1 Simulation tools VINT (Virtual InterNet Testbed): catarina.usc.edu/vint [USC/ISI, UCB,LBL,Xerox] network simulator (NS), network animator (NAM) library of protocols: • TCP variants • multicast/unicast routing • routing in ad-hoc networks • real-time protocols (RTP) • …. Other channel/protocol models & test-suites extensible framework (Tcl/tk & C++) Check the ‘Simulator’ link thru the class website Transport Layer 3-2 OPNET: commercial simulator strength in wireless channel modeling GlomoSim (QualNet): UCLA, parsec simulator Research resources: ACM & IEEE journals and conferences SIGCOMM, INFOCOM, Transactions on Networking (TON), MobiCom IEEE Computer, Spectrum, ACM Communications magazine www.acm.org, www.ieee.org Transport Layer 3-3 Modeling using queuing theory - Let: - - N be the number of sources M be the capacity of the multiplexed channel R be the source data rate be the mean fraction of time each source is active, where 0<1 Transport Layer 3-4 Transport Layer 3-5 - if N.R=M then input capacity = capacity of multiplexed link => TDM if N.R>M but .N.R<M then this may be modeled by a queuing system to analyze its performance Transport Layer 3-6 Queuing system for single server Transport Layer 3-7 is the arrival rate Tw is the waiting time The number of waiting items w=.Tw Ts is the service time is the utilization ‘fraction of the time the server is busy’, =.Ts The queuing time Tq=Tw+Ts The number of queued items (i.e. the queue occupancy) q=w+=.Tq Transport Layer 3-8 =.N.R, Ts=1/M =.Ts=.N.R.Ts=.N.R/M Assume: - random arrival process (Poisson arrival process) - constant service time (packet lengths are constant) - no drops (the buffer is large enough to hold all traffic, basically infinite) - no priorities, FIFO queue Transport Layer 3-9 Inputs/Outputs of Queuing Theory Given: - arrival rate service time queuing discipline Output: - wait time, and queuing delay - waiting items, and queued items Transport Layer 3-10 Queue Naming: X/Y/Z where X is the distribution of arrivals, Y is the distribution of the service time, Z is the number of servers G: general distribution M: negative exponential distribution (random arrival, poisson process, exponential inter-arrival time) D: deterministic arrivals (or fixed service time) Transport Layer 3-11 Transport Layer 3-12 M/D/1: Tq=Ts(2-)/[2.(1-)], q=.Tq=+2/[2.(1-)] Transport Layer 3-13 Transport Layer 3-14 Transport Layer 3-15 Transport Layer 3-16 As increases, so do buffer requirements and delay The buffer size ‘q’ only depends on Transport Layer 3-17 Queuing Example If N=10, R=100, =0.4, M=500 Or N=100, M=5000 =.N.R/M=0.8, q=2.4 - a smaller amount of buffer space per source is needed to handle larger number of sources - variance of q increases with - For a finite buffer: probability of loss increases with utilization >0.8 undesirable Transport Layer 3-18 Chapter 3 Transport Layer Computer Networking: A Top Down Approach 4th edition. Jim Kurose, Keith Ross Addison-Wesley, July 2007. Transport Layer 3-19 Transport Layer 3-20 Transport Layer 3-21 Reliable data transfer: getting started rdt_send(): called from above, (e.g., by app.). Passed data to deliver to receiver upper layer send side udt_send(): called by rdt, to transfer packet over unreliable channel to receiver deliver_data(): called by rdt to deliver data to upper receive side rdt_rcv(): called when packet arrives on rcv-side of channel Transport Layer 3-22 Flow Control - End-to-end flow and Congestion control study is complicated by: - - Heterogeneous resources (links, switches, applications) Different delays due to network dynamics Effects of background traffic We start with a simple case: hop-by-hop flow control Transport Layer 3-23 Hop-by-hop flow control Approaches/techniques for hop-by-hop flow control - Stop-and-wait sliding window - Go back N - Selective reject Transport Layer 3-24 Stop-and-wait: reliable transfer over a reliable channel underlying channel perfectly reliable no bit errors, no loss of packets stop and wait Sender sends one packet, then waits for receiver response Transport Layer 3-25 channel with bit errors underlying channel may flip bits in packet checksum to detect bit errors the question: how to recover from errors: acknowledgements (ACKs): receiver explicitly tells sender that pkt received OK negative acknowledgements (NAKs): receiver explicitly tells sender that pkt had errors sender retransmits pkt on receipt of NAK new mechanisms for: error detection receiver feedback: control msgs (ACK,NAK) rcvr->sender Transport Layer 3-26 Stop-and-wait operation Summary Stop and wait: - sender awaits for ACK to send another frame sender uses a timer to re-transmit if no ACKs if ACK is lost: - A sends frame, B’s ACK gets lost - A times out & re-transmits the frame, B receives duplicates - Sequence numbers are added (frame0,1 ACK0,1) - timeout: should be related to round trip time estimates - if too small unnecessary re-transmission - if too large long delays Transport Layer 3-27 Stop-and-wait with lost packet/frame Transport Layer 3-28 Transport Layer 3-29 Transport Layer 3-30 Stop and wait performance utilization – fraction of time sender busy sending - ideal case (error free) - u=Tframe/(Tframe+2Tprop)=1/(1+2a), a=Tprop/Tframe Transport Layer 3-31 Performance of stop-and-wait example: 1 Gbps link, 15 ms e-e prop. delay, 1KB packet: Ttransmit = L (packet length in bits) 8kb/pkt = = 8 microsec R (transmission rate, bps) 10**9 b/sec U sender: utilization – fraction of time sender busy sending U sender = L/R RTT + L / R = .008 30.008 = 0.00027 microsec onds 1KB pkt every 30 msec -> 33kB/sec thruput over 1 Gbps link network protocol limits use of physical resources! Transport Layer 3-32 stop-and-wait operation sender receiver first packet bit transmitted, t = 0 last packet bit transmitted, t = L / R first packet bit arrives last packet bit arrives, send ACK RTT ACK arrives, send next packet, t = RTT + L / R U = sender L/R RTT + L / R = .008 30.008 = 0.00027 microsec onds Transport Layer 3-33 Sliding window techniques - TCP is a variant of sliding window - Includes Go back N (GBN) and selective repeat/reject - Allows for outstanding packets without Ack - More complex than stop and wait - Need to buffer un-Ack’ed packets & more book-keeping than stop-and-wait Transport Layer 3-34 Pipelined (sliding window) protocols Pipelining: sender allows multiple, “in-flight”, yet-tobe-acknowledged pkts range of sequence numbers must be increased buffering at sender and/or receiver Two generic forms of pipelined protocols: go-Back-N, selective repeat Transport Layer 3-35 Pipelining: increased utilization sender receiver first packet bit transmitted, t = 0 last bit transmitted, t = L / R first packet bit arrives last packet bit arrives, send ACK last bit of 2nd packet arrives, send ACK last bit of 3rd packet arrives, send ACK RTT ACK arrives, send next packet, t = RTT + L / R Increase utilization by a factor of 3! U sender = 3*L/R RTT + L / R = .024 30.008 = 0.0008 microsecon ds Transport Layer 3-36 Go-Back-N Sender: k-bit seq # in pkt header “window” of up to N, consecutive unack’ed pkts allowed ACK(n): ACKs all pkts up to, including seq # n - “cumulative ACK” may receive duplicate ACKs (more later…) timer for each in-flight pkt timeout(n): retransmit pkt n and all higher seq # pkts in window Transport Layer 3-37 GBN: receiver side ACK-only: always send ACK for correctly-received pkt with highest in-order seq # may generate duplicate ACKs need only remember expected seq num out-of-order pkt: discard (don’t buffer) -> no receiver buffering! Re-ACK pkt with highest in-order seq # Transport Layer 3-38 GBN in action Transport Layer 3-39 Selective Repeat receiver individually acknowledges all correctly received pkts buffers pkts, as needed, for eventual in-order delivery to upper layer sender only resends pkts for which ACK not received sender timer for each unACKed pkt sender window N consecutive seq #’s limits seq #s of sent, unACKed pkts Transport Layer 3-40 Selective repeat: sender, receiver windows Transport Layer 3-41 Selective repeat in action Transport Layer 3-42 performance: - selective repeat: - error-free case: - if the window is w such that the pipe is fullU=100% - otherwise U=w*Ustop-and-wait=w/(1+2a) - in case of error: - if w fills the pipe U=1-p - otherwise U=w*Ustop-and-wait=w(1-p)/(1+2a) Transport Layer 3-43 TCP: Overview point-to-point: one sender, one receiver reliable, in-order byte stream: no “message boundaries” pipelined: TCP congestion and flow control set window size send & receive buffers socket door application writes data application reads data TCP send buffer TCP receive buffer RFCs: 793, 1122, 1323, 2018, 2581 full duplex data: bi-directional data flow in same connection MSS: maximum segment size connection-oriented: handshaking (exchange of control msgs) init’s sender, receiver state before data exchange flow controlled: sender will not socket door overwhelm receiver segment Transport Layer 3-44 TCP segment structure 32 bits URG: urgent data (generally not used) ACK: ACK # valid PSH: push data now (generally not used) RST, SYN, FIN: connection estab (setup, teardown commands) Internet checksum (as in UDP) source port # dest port # sequence number acknowledgement number head not UA P R S F len used checksum Receive window Urg data pnter Options (variable length) counting by bytes of data (not segments!) # bytes rcvr willing to accept application data (variable length) Transport Layer 3-45 - Receive window: credit (in octets) that the receiver is willing to accept from the sender starting from ack # - flags: - SYN: synchronizing at initail connection time FIN: end of sender data PSH: when used at sender the data is transmitted immediately, when at receiver, it is accepted immediately - options: - window scale factor (WSF): actual window = 2Fxwindow field, where F is the number in the WSF - timestamp option: helps in RTT (round-trip-time) calculations Transport Layer 3-46 credit allocation scheme - - (A=i,W=j) [A=Ack, W=window]: receiver acks up to ‘i-1’ bytes and allows/anticipates i up to i+j-1 receiver can use the cumulative ack option and not respond immediately - performance: depends on - transmission rate, propagation, window size, queuing delays, retransmission strategy which depends on RTT estimates that affect timeouts and are affected by network dynamics, receive policy (ack), background traffic….. it is complex! Transport Layer 3-47 TCP seq. #’s and ACKs Seq. #’s: byte stream “number” of first byte in segment’s data ACKs: seq # of next byte expected from other side cumulative ACK Q: how receiver handles out-of-order segments A: TCP spec doesn’t say, - up to implementor Host A User types ‘C’ Host B host ACKs receipt of ‘C’, echoes back ‘C’ host ACKs receipt of echoed ‘C’ simple telnet scenario time Transport Layer 3-48 TCP retransmission strategy: - TCP performs end-to-end flow/congestion control and error recovery - TCP depends on implicit congestion signaling and uses an adaptive retransmission timer, based on average observation of the ack delays. Transport Layer 3-49 - Ack delays may be misleading due to the following reasons: - - Cumulative acks render this estimate inaccurate Abrupt changes in the network If ack is received for a re-transmitted packet, sender cannot distinguish between ack for the original packet and ack for the re-transmitted packet Transport Layer 3-50 Reliability in TCP Components of reliability 1. Sequence numbers 2. Retransmissions 3. Timeout Mechanism(s): function of the round trip time (RTT) between the two hosts (is it static?) Transport Layer 3-51 TCP Round Trip Time and Timeout Q: how to set TCP timeout value? longer than RTT but RTT varies too short: premature timeout unnecessary retransmissions too long: slow reaction to segment loss Q: how to estimate RTT? SampleRTT: measured time from segment transmission until ACK receipt ignore retransmissions SampleRTT will vary, want estimated RTT “smoother” average several recent measurements, not just current SampleRTT Transport Layer 3-52 TCP Round Trip Time and Timeout EstimatedRTT(k) = (1- )*EstimatedRTT(k-1) + *SampleRTT(k) =(1- )*((1- )*EstimatedRTT(k-2)+ *SampleRTT(k-1))+ *SampleRTT(k) =(1- )k *SampleRTT(0)+ (1- )k-1 *SampleRTT)(1)+…+ *SampleRTT(k) Exponential weighted moving average (EWMA) influence of past sample decreases exponentially fast typical value: = 0.125 Transport Layer 3-53 Example RTT estimation: RTT: gaia.cs.umass.edu to fantasia.eurecom.fr 350 RTT (milliseconds) 300 250 200 150 100 1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 time (seconnds) SampleRTT Estimated RTT Transport Layer 3-54 =0.5 =0.125 Transport Layer 3-55 =0.125 =0.125 Transport Layer 3-56 TCP Round Trip Time and Timeout Setting the timeout EstimtedRTT plus “safety margin” large variation in EstimatedRTT -> larger safety margin 1. estimate how much SampleRTT deviates from EstimatedRTT: DevRTT = (1-)*DevRTT + *|SampleRTT-EstimatedRTT| (typically, = 0.25) 2. set timeout interval: TimeoutInterval = EstimatedRTT + 4*DevRTT 3. For further re-transmissions (if the 1st re-tx was not Ack’ed) - RTO=q.RTO, q=2 for exponential backoff - similar to Ethernet CSMA/CD backoff Transport Layer 3-57 TCP reliable data transfer TCP creates reliable service on top of IP’s unreliable service Pipelined segments Cumulative acks TCP uses single retransmission timer Retransmissions are triggered by: timeout events duplicate acks Initially consider simplified TCP sender: ignore duplicate acks ignore flow control, congestion control Transport Layer 3-58 TCP: retransmission scenarios Host A X loss Sendbase = 100 SendBase = 120 SendBase = 100 time SendBase = 120 lost ACK scenario Host B Seq=92 timeout Host B Seq=92 timeout timeout Host A time premature timeout Transport Layer 3-59 TCP retransmission scenarios (more) timeout Host A Host B X loss SendBase = 120 time Cumulative ACK scenario Transport Layer 3-60 Fast Retransmit Time-out period often relatively long: long delay before resending lost packet Detect lost segments via duplicate ACKs. Sender often sends many segments back-toback If segment is lost, there will likely be many duplicate ACKs. If sender receives 3 ACKs for the same data, it supposes that segment after ACKed data was lost: fast retransmit: resend segment before timer expires Transport Layer 3-61 (Self-clocking) Transport Layer 3-62 TCP Flow Control receive side of TCP connection has a receive buffer: flow control sender won’t overflow receiver’s buffer by transmitting too much, too fast match the send rate to the receiving app’s drain rate app process may be slow at reading from buffer (low drain rate) Transport Layer 3-63 Principles of Congestion Control Congestion: informally: “too many sources sending too much data too fast for network to handle” different from flow control! manifestations: lost packets (buffer overflow at routers) long delays (queueing in router buffers) a key problem in the design of computer networks Transport Layer 3-64 Congestion Control & Traffic Management - Does adding bandwidth to the network or increasing the buffer sizes solve the problem of congestion? No. We cannot over-engineer the whole network due to: -Increased traffic from applications (multimedia,etc.) -Legacy systems (expensive to update) -Unpredictable traffic mix inside the network: where is the bottleneck? Congestion control & traffic management is needed To provide fairness To provide QoS and priorities Transport Layer 3-65 Network Congestion - Modeling the network as network of queues: (in switches and routers) - Store and forward Statistical multiplexing Limitations: -on buffer size -> contributes to packet loss - if we increase buffer size? - excessive delays - if infinite buffers - infinite delays Transport Layer 3-66 - solutions: - policies for packet service and packet discard to limit delays congestion notification and flow/congestion control to limit arrival rate buffer management: input buffers, output buffers, shared buffers Transport Layer 3-67 Notes on congestion and delay - fluid flow model Arrival Rate Departure Rate - arrival > departure --> queue build-up --> overflow and excessive delays - TTL field: time-to-live - Limits number of hops traversed Limits the time - Infinite buffer --> queue build-up and TTL decremented --> Tput goes to 0 Transport Layer 3-68 Using the fluid flow model to reason about relative flow delays in the Internet Service Time: Ts=1/BWoutput Flow Arrival BWinput Bwoutput - Bandwidth is split between flows such that flow 1 gets f1 fraction, flow 2 gets f2 … so on. Transport Layer 3-69 f1 is fraction of the bandwidth given to flow 1 f2 is fraction of the bandwidth given to flow 2 1 is the arrival rate for flow 1 2 is the arrival rate for flow 2 for M/D/1: delay Tq=Ts[1+/[2(1-)]] The total server utilization, =Ts. Fraction time utilized by flow i, Ti =Ts/fi (or the bandwidth utilized by flow i, Bi=Bs.fi, where Bi=1/Ti and Bs=1/Ts=M [the total b.w.]) The utilization for flow i, i = i.Ti= i/(Bs.fi) Transport Layer 3-70 Tq and q = f() If utilization is the same, then queuing delay is the same Delay for flow i= f(i) i= i.Ti= Ts.i/fi Condition for constant delay for all flows i/fi is constant Transport Layer 3-71 Propagation of congestion - if flow control is used hop-by-hop then congestion may propagate throughout the network Transport Layer 3-72 congestion phases and effects - ideal case: infinite buffers, - Tput increases with demand & saturates at network capacity Tput/Gput Delay Network Power = Tput/delay Transport Layer Representative of Tput-delay design trade-off 3-73 practical case: finite buffers, loss - no congestion --> near ideal performance - overall moderate congestion: - severe congestion in some nodes - dynamics of the network/routing and overhead of protocol adaptation decreases the network Tput - severe congestion: - loss of packets and increased discards - extended delays leading to timeouts - both factors trigger re-transmissions - leads to chain-reaction bringing the Tput down Transport Layer 3-74 Normalized Goodput Network Congestion Phases (I) (II) (III) Load (I) No Congestion (II) Moderate Congestion (III) Severe Congestion (Collapse) What is the best operational point and how do we get (and stay) there? Transport Layer 3-75 Congestion Control (CC) - Congestion is a key issue in network design - various techniques for CC 1.Back pressure - hop-by-hop flow control (X.25, HDLC, Go back N) - May propagate congestion in the network 2.Choke packet - generated by the congested node & sent back to source - example: ICMP source quench - sent due to packet discard or in anticipation of congestion Transport Layer 3-76 Congestion Control (CC) (contd.) 3.Implicit congestion signaling - - used in TCP delay increase or packet discard to detect congestion may erroneously signal congestion (i.e., not always reliable) [e.g., over wireless links] done end-to-end without network assistance TCP cuts down its window/rate Transport Layer 3-77 Congestion Control (CC) (contd.) 4.Explicit congestion signaling - (network assisted congestion control) gets indication from the network - forward: going to destination - backward: going to source - 3 approaches - Binary: uses 1 bit (DECbit, TCP/IP ECN, ATM) - Rate based: specifying bps (ATM) - Credit based: indicates how much the source can send (in a window) Transport Layer 3-78 Transport Layer 3-79 TCP congestion control: additive increase, multiplicative decrease Approach: increase transmission rate (window size), probing for usable bandwidth, until loss occurs additive increase: increase rate (or congestion window) CongWin until loss detected multiplicative decrease: cut CongWin in half after loss Saw tooth behavior: probing for bandwidth congestion window size congestion window 24 Kbytes 16 Kbytes 8 Kbytes timetime Transport Layer 3-80 TCP Congestion Control: details sender limits transmission: LastByteSent-LastByteAcked CongWin Roughly, rate = CongWin Bytes/sec RTT CongWin is dynamic, function of perceived network congestion How does sender perceive congestion? loss event = timeout or duplicate Acks TCP sender reduces rate (CongWin) after loss event three mechanisms: AIMD slow start conservative after timeout events Transport Layer 3-81 TCP window management - At any time the allowed window (awnd): awnd=MIN[RcvWin, CongWin], - where RcvWin is given by the receiver (i.e., Receive Window) and CongWin is the congestion window - Slow-start algorithm: - start with CongWin=1, then CongWin=CongWin+1 with every ‘Ack’ This leads to ‘doubling’ of the CongWin with RTT; i.e., exponential increase Transport Layer 3-82 TCP Slow Start (more) When connection Host B RTT begins, increase rate exponentially until first loss event: Host A double CongWin every RTT done by incrementing CongWin for every ACK received Summary: initial rate is slow but ramps up exponentially fast time Transport Layer 3-83 TCP congestion control Initially we use Slow start: CongWin = CongWin + 1 with every Ack When timeout occurs we enter congestion avoidance: - ssthresh=CongWin/2, CongWin=1 slow start until ssthresh, then increase ‘linearly’ CongWin=CongWin+1 with every RTT, or CongWin=CongWin+1/CongWin for every Ack - additive increase, multiplicative decrease (AIMD) Transport Layer 3-84 Transport Layer 3-85 Congestion Avoidance Linear increase CongWin Slow start Exponential increase (RTT) Transport Layer 3-86 Fast Retransmit & Recovery Fast retransmit: - receiver sends Ack with last in-order segment for every out-of-order segment received when sender receives 3 duplicate Acks it retransmits the missing/expected segment Fast recovery: when 3rd dup Ack arrives - ssthresh=CongWin/2 - retransmit segment, set CongWin=ssthresh+3 CongWin - for every duplicate Ack: CongWin=CongWin+1 (note: beginning of window is ‘frozen’) - after receiver gets cumulative Ack: CongWin=ssthresh (beginning of window advances to last Ack’ed segment) Transport Layer 3-87 Transport Layer 3-88 TCP Fairness Fairness goal: if K TCP sessions share same bottleneck link of bandwidth R, each should have average rate of R/K TCP connection 1 TCP connection 2 bottleneck router capacity R Transport Layer 3-89 Fairness (more) Fairness and UDP Multimedia apps often do not use TCP do not want rate throttled by congestion control Instead use UDP: pump audio/video at constant rate, tolerate packet loss Research area: TCP friendly protocols! Fairness and parallel TCP connections nothing prevents app from opening parallel connections between 2 hosts. Web browsers do this Example: link of rate R supporting 9 connections; new app asks for 1 TCP, gets rate R/10 new app asks for 11 TCPs, gets R/2 ! Transport Layer 3-90 Congestion Control with Explicit Notification - TCP uses implicit signaling - ATM (ABR) uses explicit signaling using RM (resource management) cells - ATM: Asynchronous Transfer Mode, ABR: Available Bit Rate ABR Congestion notification and congestion avoidance - parameters: - peak cell rate (PCR) minimum cell rate (MCR) initial cell rate(ICR) Transport Layer 3-91 - ABR uses resource management cell (RM cell) with fields: - - CI (congestion indication) NI (no increase) ER (explicit rate) Types of RM cells: - Forward RM (FRM) - Backward RM (BRM) Transport Layer 3-92 Transport Layer 3-93 Congestion Control in ABR - The source reacts to congestion notification by decreasing its rate (ratebased vs. window-based for TCP) - Rate adaptation algorithm: - If CI=0,NI=0 - Rate increase by factor ‘RIF’ (e.g., 1/16) - Rate = Rate + PCR/16 - Else If CI=1 - Rate decrease by factor ‘RDF’ (e.g., 1/4) - Rate=Rate-Rate*1/4 Transport Layer 3-94 Transport Layer 3-95 Which VC to notify when congestion occurs? - - FIFO, if Qlength > 80%, then keep notifying arriving cells until Qlength < lower threshold (this is unfair) Use several queues: called Fair Queuing Use fair allocation = target rate/# of VCs = R/N - If current cell rate (CCR) > fair share, then notify the corresponding VC Transport Layer 3-96 What to notify? CI NI ER (explicit rate) schemes perform the steps: – Compute the fair share – Determine load & congestion – Compute the explicit rate & send it back to the source Should we put this functionality in the network? Transport Layer 3-97